DeepSeek and Huawei to open-source Ascend 950 tooling, pushing TileLang as a CUDA alternative

This partnership targets NVIDIA's strongest advantage, its CUDA software ecosystem, rather than its chips. The New York Times reports that DeepSeek and Huawei will jointly build and open-source two things:

- compute and communication libraries for Huawei's Ascend 950 accelerators

- a reference design for a 128-chip supernode

The programming centerpiece is TileLang. It is a high-level language designed to be simpler than writing CUDA-style kernels, while still reaching peak performance on Ascend hardware. Quartz described the release as open-source AI chip software aimed squarely at NVIDIA.

The compute commitment is large. DeepSeek plans to deploy more than 160,000 Huawei accelerators in Inner Mongolia. Combined with open-source tooling, that gives Chinese labs a credible path to train and serve frontier-scale models without NVIDIA's stack. US export controls make that path both necessary and politically significant.

It also lands as DeepSeek's models are winning on volume. DeepSeek V4.1 Flash led OpenRouter's weekly usage for a second week, and Alibaba is touting progress on its own models and chips. China's AI ecosystem is assembling its own hardware, compilers and cheap inference.

Community reaction is strongly interested. Developers see TileLang as the most credible challenge to CUDA so far, mainly because a simpler programming model makes migration easier. Caveats:

- No timeline or license details were reported.

- Ascend 950 performance relative to NVIDIA Blackwell-class parts is unverified in these sources.

- Developer ecosystems take years to mature. CUDA's lead comes from libraries, tooling and talent, not just syntax.